The daily brief on robots that learn from video
Six sources cut overnight to the ~30 that matter: a PDF at 8:30. A real issue is right below. Scroll it.
The real issue of August 20, 2026, built from ~1,600 items in one night of pipeline data. Read it first: if it does not hold your attention for ten minutes, you have your answer.
Six sources, ranked overnight, merged into one issue
Founding price: $9/month
- $9/mo locked for as long as you stay. The price rises for new subscribers after the first 100. I haven't set the new number, but your rate stays $9 for as long as you keep the subscription
- A new issue emailed to you every morning, at 8:30 your local time
- All six sources, research to discourse, as a print-ready PDF, yours to keep
- Every item links to the paper, post or thread
Who makes this
I'm Siddhesh. I studied engineering at IIT Gandhinagar and build RL training environments for frontier AI labs at Deeptune (acq. by Mercor). Following this niche on X and LinkedIn meant so much noise and promoted content that I eventually gave up on tracking the field, so I built Pixels2Actions. I read every digest with you.
Questions
Who is this for?
Researchers and engineers working on robot learning, imitation learning, VLAs and world models, plus people working their way into the field and the founders and investors tracking physical AI. It is written for people who want the depth: the issues are dense by design, It can take upto 15 to 20 involved minutes.
How does it work?
"Physical AI" is too broad to track as one query, so I split it into six micro-niches: learning from human video, latent actions and world models, video models as policies and simulators, VLA foundation models, teleoperation and data engines, and rewards and evaluation. Each niche is tracked separately, with its own sources and its own relevance profile. The run is nightly, in three stages. Sweep: the full arXiv announcement batch plus the ranked top of the other five sources. On a typical night this is about 1,600 items. Score: items are scored for topical relevance against each niche's profile. The niche results are merged, deduped, and cut to the ~30 highest-scoring. Typeset: the remaining items are laid out as a broadsheet: front page, research, newsletters, the discourse. A slow news day produces a thinner issue. The pipeline does not pad.
Is this AI-generated slop?
No language model searches the web, picks the stories, or writes the news. The pipeline has three stages: search pulls ~1,600 items a night from six sources, dedup collapses reposts and mirrors, and selection ranks what's left by topical relevance. A language model is used in one place, after selection: tidying arXiv abstracts and X posts so they read cleanly; titles and summaries otherwise stay in the source's own words. Every item links to the original, and each issue's footer prints the keywords its ranking optimized for, so you can audit both.
How is it different from Weekly Robotics or The Robot Report?
They're weekly and cover all of robotics; this is daily and narrow: how robots learn from video, spanning human demonstrations, latent actions and world models (JEPA and beyond), video-generation policies, VLAs, robot data engines, and policy evaluation.
What's in the pipeline?
Three things. More sources: company-specific tracking (following the ~100 companies that matter in this field directly), plus YouTube, Hugging Face and GitHub. Better selection: today's digests still let some noise through, anywhere from 5–25% of an issue depending on the night, and reducing that is ongoing work. More niches: more corners of physical AI first, then other fields on demand.
How do I cancel? What about refunds?
Two options: cancel the subscription from your PayPal account, or email siddhesh@pixelstoactions.com and I'll cancel it for you. I refund the current month on request, no questions asked.
The pipeline runs every night regardless; subscribing puts the result in your inbox at 8:30.